Deep Learning Methods for Magnetic Resonance Spectroscopic Neuroimaging
Author(s)
Giuffrida, Alexander
Advisor(s)
Editor(s)
Collections
Supplementary to:
Permanent Link
Abstract
Spectroscopic magnetic resonance imaging (sMRI) enables noninvasive mapping of brain metabolism, but its clinical use has been limited by challenges in spectral quantification, quality control, and interpretation. This dissertation introduces deep learning methods and software tools to address these limitations and improve the speed, scalability, and reliability of sMRI analysis. The central contribution is NNFit, a self-supervised neural network for spectral fitting that replaces slow, iterative optimization with a single forward pass. NNFit reduces scan processing time from nearly an hour to under 15 seconds while maintaining strong agreement with conventional methods. The model was validated on both short- and long-echo time data and applied to treatment planning in glioblastoma patients, demonstrating its potential for real-world clinical workflows. To support interpretation, the Onix software platform was developed for visualizing spectra, comparing methods, and computing regional metrics. A ResNet-based artifact filtration model was also integrated into Onix, improving quality control and reproducibility across the pipeline. Together, these contributions provide a foundation for faster, more interpretable, and clinically relevant sMRI workflows. The tools and models developed in this work can be extended to other diseases and imaging protocols, and offer a flexible foundation for future innovations in metabolic neuroimaging and AI-assisted clinical decision-making.
Sponsor
Date
2025-07-22
Extent
Resource Type
Text
Resource Subtype
Dissertation